Automatic high frequency ultrasound image segmentation and shape analysis.

摘要

High frequency ultrasound (HFU) is widely used in imaging biological tissues, but it suffers from low contrast, speckle noise, and acoustic attenuation. In this thesis, we focus on three dimensional (3D) HFU image segmentation and segmentation result analysis. Toward these goals, algorithms for image segmentation, spatially varying intensity distribution statistics estimation, and 3D shape decomposition and characterization are proposed.;We firstly propose a fully automatic segmentation method called nested graph cut (NGC) to segment images (2D or 3D) that contain multiple objects with a nested structure. Compared to other graph-cut-based methods developed for multiple regions, our method can work well for nested objects without requiring manual selection of initial seeds, even if different objects have similar intensity distributions and some object boundaries are missing. Promising results were obtained for separating the brain ventricles (BVs), the head, and the uterus region in the 3D mouse-embryo head images obtained using HFU imaging.;Next we consider the segmentation of lymph nodes (LNs) in HFU human LN images, which contain three different parts: LN-parenchyma (LNP), fat, and phosphate-buffered saline (PBS). The main challenge is the large spatial variability of the spatially varying statistics of intensity distribution of LNP and fat pixels due to acoustic attenuation and focusing. To overcome this issue, we proposed two methods to estimate the intensity profile of LNP and fat. The first one is an iterative self-updating segmentation framework combining NGC and robust spline fitting to estimate the depth-dependent intensity mean and variance. The second approach estimates three smooth 3D intensity mean profiles for LNP, fat, and PBS, respectively from a given HFU image, using a random sample consensus (RANSAC) like robust regression method. Compared to depth dependent profiles, 3D spatially varying intensity profiles can model the variability of intensity distributions in all directions. By using these estimated intensity distributions to determine the energy term, NGC can segment LNs in HFU images accurately even when the acoustic attenuation is strong and highly inhomogeneous.;Finally, we explore volumetric analysis of BVs of mouse embryos, which is important to the study of normal and abnormal development of the central nervous system of mouse embryos. Specifically, we develop methods for automatic staging and mutant detection from the segmented BV shapes from the HFU images. We present novel algorithms for deriving the Y-skeleton representation of a BV and decomposing the BV volume into five components (fourth ventricle, aqueduct, third ventricle and two lateral ventricles). Embryo staging and mutant detection are accomplished by analyzing the volume profile along the Y-skeleton and the volumes of the five BV components.